Anthropic Tunes AI Chatbot with Community Feedback to Reduce Biased Outputs

Anthropic developed an AI chatbot by surveying 1,000 users on its values via Polis. The results were used to fine-tune a large language model.

The goal was to make the AI more democratic by letting users determine appropriate values and outputs.

Anthropic used a method called Constitutional AI, giving the model rules to follow based on user feedback.

The tuned model slightly outperformed the base model in reducing biased outputs.

Anthropic hopes this community-directed training method can allow groups to develop models suited to their cultural needs.